Identification of cervical cancer stem cells using single-cell transcriptomes of normal cervix, cervical premalignant lesions, and cervical cancer

Summary

Background

Cervical cancer is the fourth leading cause of mortality among gynecological malignancies. However, the identification of cervical cancer stem cells remains unclear.

Methods

We performed single-cell mRNA sequencing on ∼122,400 cells from 20 cervical biopsies, including 5 healthy controls, 4 high-grade intraepithelial neoplasias, 5 microinvasive carcinomas of the cervix, and 6 invasive cervical squamous carcinomas. Bioinformatic results were validated by multiplex immunohistochemistry (mIHC) in cervical cancer tissue microarrays (TMA) (n = 85).

Findings

We identified cervical cancer stem cells and highlighted the functional changes in cervical stem cells during malignant transformation. The original non-malignant stem cell properties (characterized by high proliferation) gradually diminished, whereas the tumor stem cell properties (characterized by epithelial-mesenchymal transformation and invasion) were enhanced. The mIHC results of our TMA cohort confirmed the existence of stem-like cells and indicated that cluster correlated with neoplastic recurrence. Subsequently, we investigated malignant and immune cell heterogeneity in the cervical multicellular ecosystem across different disease stages. We observed global upregulation of interferon responses in the cervical microenvironment during lesion progression.

Interpretation

Our results provide more insights into cervical premalignant and malignant lesion microenvironments.

Funding

This research was supported by the Guangdong Provincial Natural Science Foundation of China (2023A1515010382), Grant 2021YFC2700603 from the National Key Research & Development Program of China and the Hubei Provincial Natural Science Foundation of China (2022CFB174 and 2022CFB893).

Research in context.

Evidence before this study

Cervical cancer develops from a well-defined precancerous phase. These premalignant changes can be diagnosed via cytological examination and pathological analysis of cervical tissue. Epidemiological investigations reveal that most females are infected with HPV at some time in life; however, only a few develop invasive cervical cancer. Current screening and examinations cannot accurately distinguish between the minority who may progress to invasive cancer and the majority whose abnormalities will regress spontaneously.

This inaccuracy is largely because of our limited knowledge of the cellular and molecular mechanisms underlying the natural progression of cervical intraepithelial neoplasia, despite its clear histological definition. In addition, the cervical cancer stem cells or tumor-initiating cells also remain unclear.

Added value of this study

This study focused on identifying cervical cancer stem cells and compared cervical ecosystems across different disease stages using single-cell mRNA sequencing. We identified cervical cancer stem cells and highlighted the functional changes in cervical stem cells during malignant transformation. The original non-malignant stem cell properties (characterized by high proliferation) gradually diminished, whereas the tumor stem cell properties (characterized by epithelial-mesenchymal transformation and invasion) were enhanced. The mIHC results of our TMA cohort confirmed the existence of stem-like cells and indicated that cluster correlated with neoplastic recurrence.

Implications of all the available evidence

This study explored the malignant transformation process of cervical stem cells to tumor stem cells, which deepened the understanding of the mechanism of cervical carcinogenesis. We provided theoretical implications for the understanding and treatment of tumor behaviors such as drug resistance and metastasis associated with cervical tumor stem cells.

Introduction

Despite effective cervical cancer screening programs and mass vaccination campaigns, cervical cancer remains a major global health concern. Cervical cancer is the fourth most common malignancy in women worldwide, with an estimated 569,847 new cases and 311,365 deaths in 2018.¹ To date, treatment options for advanced cervical cancer are limited. Lesion recurrence and drug-resistant tumors have long been the focus of research.^2,3 Human papillomavirus (HPV) belongs to the papillomavirus family, which includes over 100 virus subtypes. Persistent high-risk HPV infections cause almost all cervical cancer cases.^1,4 Transforming protein E6 (E6) and transforming protein E7 (E7), which are encoded by high-risk HPV, effectively promote cellular proliferation and prevent apoptosis. The resulting genomic instability causes genetic alteration accumulation, ultimately leading to malignant transformation of HPV-infected cells.⁵

Cervical cancer develops from a well-defined precancerous phase manifesting histological abnormalities ranging from low-grade squamous intraepithelial lesions (LSILs) to high-grade squamous intraepithelial lesions (HSILs).⁵ These premalignant changes can be diagnosed via cytological examination and pathological analysis of cervical tissue.⁶ Epidemiological investigations reveal that most females are infected with HPV at some time in life; however, only a few develop invasive cervical cancer.⁵ Current screening and examinations cannot accurately distinguish between the minority who may progress to invasive cancer and the majority whose abnormalities will regress spontaneously. This inaccuracy is largely because of our limited knowledge of the cellular and molecular mechanisms underlying the natural progression of cervical intraepithelial neoplasia, despite its clear histological definition. In addition, the key targets of HPV infection, that is, cervical cancer stem cells (CSCs) or tumor-initiating cells, also remain unclear.^7,8

Thus, comprehensively understanding the multicellular ecosystem of cervical neoplasia is critical. Single-cell mRNA sequencing (scRNA-seq) is a powerful tool to gain deeper insights into cervical microenvironments. In this study, we conducted scRNA-seq on ∼122,400 cells from 20 cervical biopsies from five healthy controls, four HSIL patients, five patients with microinvasive carcinomas of the cervix (MIC), and six invasive cervical squamous carcinoma (ICC) patients. We focused on identifying cervical CSCs and compared cervical ecosystems across different disease stages.

Methods

Patients and sample collection

Fresh cervix tissue samples were collected from cases undergoing cervical biopsy procedure or surgical in Tongji Hospital of Tongji Medicine College of Huazhong University of Science and Technology (HUST). 20 cervical biopsies were collected, including 5 healthy control (HPV−), 4 high-grade intraepithelial neoplasia (HPV+), 5 microinvasive carcinoma of the cervix (HPV+) and 6 invasive cervical squamous carcinoma (HPV+). The control group had a normal cervix and they underwent hysterectomy for symptomatic hysteromyoma (Supplementary information, Table S1).

Preparation of cell suspensions

scRNA-seq experiment was performed by experimental personnel in the laboratory of GENECHEM. The tissues were surgically removed and kept in MACS Tissue Storage Solution (Miltenyi Biotec) until processing. The tissue samples were processed as described below. Briefly, samples were first washed with phosphate-buffered saline (PBS), minced into small pieces (approximately 1mm3) on ice and enzymatically digested with 1 mg/mL collagenase I (Worthington) and 15 U/mL DNase I (Worthington) for 45 min at 37 °C, with agitation. After digestion, samples were sieved through a 70 μm cell strainer, and centrifuged at 300 g for 5 min. After the supernatant was removed, the pelleted cells were suspended in red blood cell lysis buffer (Miltenyi Biotec) to lyse red blood cells. After washing with PBS containing 0.04% BSA, the cell pellets were re-suspended in PBS containing 0.04% BSA and re-filtered through a 40 μm cell strainer. Dissociated single cells were then stained for viability assessment using Calcein-AM (Thermo Fisher Scientific) and Draq7 (BD Biosciences).

scRNA-seq

BD Rhapsody system was used to capture the transcriptomic information of the single cells. Single-cell capture was achieved by random distribution of a single-cell suspension across >200,000 microwells through a limited dilution approach. Beads with oligonucleotide barcodes were added to saturation so that a bead was paired with a cell in a microwell. The cells were lysed in the microwell to hybridize mRNA molecules to barcoded capture oligos on the beads. Beads were collected into a single tube for reverse transcription and ExoI digestion. Upon cDNA synthesis, each cDNA molecule was tagged on the 5′ end (that is, the 3′ end of a mRNA transcript) with a unique molecular identifier (UMI) and cell barcode indicating its cell of origin. Whole transcriptome libraries were prepared using the BD Rhapsody single-cell whole-transcriptome amplification (WTA) workflow including random priming and extension (RPE), RPE amplification PCR and WTA index PCR. The libraries were quantified using a High Sensitivity DNA chip (Agilent) on a Bioanalyzer 2200 and the Qubit High Sensitivity DNA assay (Thermo Fisher Scientific). Sequencing was performed by illumina sequencer (Illumina, San Diego, CA) on a 150 bp paired-end run.

Alignment, quantification and quality control of scRNA-seq data

BD scRNA-seq data was quantified using the BD Single-Cell analysis Software Suite (Version 1.8) and aligned to the GRCh38 reference genome. Cells with fewer than 200 genes or more than 6000 genes detected, and for which the total mitochondrial gene expression exceeded 50% were removed. Genes that were expressed in fewer than 2 cells were also removed.

Data embedding, visualization, clustering and annotation

Principal components were calculated using scanpy.pca. Dimensionality reduction and embedding was performed using Uniform Manifold Approximation and Projection (UMAP) by the scanpy.tl.umap function. Differentially expressed genes were calculated using the Wilcoxon sum rank test restricted to genes expressed in at least 30% of cells in either of the two populations compared, and with a fold change cutoff of 2. All p-values were adjusted for multiple testing using the Bonferroni correction. Initial annotation was ascribed by comparing these differentially expressed genes to published bulk transcription profiles or protein expression of defined cell types (singleR).

Single cell RNA CNV evaluation

The CNV evaluation based on scRNA-seq raw counts was conducted by infercnv R package. We chose hidden Markov model to predict the CNV states.

Cell stemness analysis

Stemness index (si) based on mRNA expression were calculated using one-class logistic regression (OCLR) as the previous study.⁹

EMT score analysis

EMT score based on mRNA expression were calculated as the previous study.¹⁰ Briefly, four established EMT markers were used as seeds to derive the pan-cancer EMT signature on the basis of TCGA pan-cancer RNAseq data, including CDH1 (epithelial marker, E type), CDH2 (mesenchymal marker, M type), VIM (M type) and FN1 (M type).

Cell cycle analysis

The cell cycle status of both malignant and nonmalignant cells were calculated by scanpy.score_genes_cell_cycle based on a previously defined gene list containing 43 G1/S genes and 54 G2/M genes.

Trajectory analysis

Reconstruction of cell differentiation pathway was performed using partition-based approximate graph abstraction (PAGA) in Scanpy v1.4 (sc.tl.paga) with default parameters. PAGA represents clusters with nodes, and edge weights (thickness) quantify the connectivity between groups. Connectivity is a measure of the fraction of inter-cluster edges in the full neighborhood graph of single cells.

HPV-derived sequencing reads analysis

We applied BD Rhapsody Analysis pipeline (Version 1.8 https://hub.docker.com/r/bdgenomics/rhapsody) with default parameter to align reads and generate feature-barcode matrices. The reads were mapped to human genome (GENCODE version 29) and high-risk HPV genome (https://pave.niaid.nih.gov/#home) (including HPV16, HPV18, HPV31, HPV33, HPV35, HPV39, HPV45, HPV51, HPV52, HPV56, HPV58, HPV59, HPV68, HPV26, HPV30, HPV34, HPV53, HPV66, HPV67, HPV69, HPV70, HPV73, HPV82, HPV85, HPV97) utilizing STAR mapping with customized parameter from BD Rhapsody standard pipeline.

Cell–cell communication analysis

To assess cell–cell communication molecules between different cell types, we used CellPhoneDB software (version 1.1.0) (www.cellphonedb.org) to infer the intercellular communication network from single-cell transcriptome data. First, by randomly permuting the cluster labels of all cells 1000 times, we determined the mean of the average receptor expression level of a cluster and the average ligand expression level of the interacting cluster. Then, we performed pairwise comparisons between all celltypes and obtained a likelihood of p-value to filter the false–positive interaction. Only interactions with p < 0.05 are considered to be real.

Deciphering intratumor expression programs and meta-programs of malignant cell

We used a non-negative factorization algorithm (documented by the R NMF package) to identify underlying intratumor expression programs within the 15 samples of HSIL, MIC and ICC groups. For each sample, we applied NMF to the relative expression matrix with all negative values replaced by zero. We filtered genes with standard deviations of expression <0.5 within each sample. We successfully extracted 10 potential programs from each tumor, which generated 150 intratumor expression programs among the 15 samples. To investigate common programs among the 150 signatures, we calculated the Pearson correlation coefficient of these programs, based on which they were clustered into meta-programs. Nine prominent meta-programs were identified and further analyzed.

Expression programs of intratumor heterogeneity and meta-signatures

A non-negative factorization algorithm (documented by the R NMF package) was used to identify underlying intratumor expression programs within the 10 tumors with the largest numbers of malignant cells. For each tumor, we applied NMF to the relative expression matrix with all negative values replaced by zero. A total of 5 potential programs from each tumor, which generated 150 intratumor expression programs among the 15 samples.

Gene set score analysis

Pathway enrichment analyses were performed using the ssGSEA function from gseapy package with default parameters Gene sets were obtained from the MSigDB database (https://www.gsea-msigdb.org/gsea/msigdb/).

Cohort and immunohistochemistry

The cervical cancer tissue microarrays, purchased from Shanghai Outdo Biotech (Shanghai, China), contained 85 cases of cancerous tissues with follow-up and 22 matiched cases of paracancerous tissues. This process had fully informed consent of the patients. Immunohistochemistry was performed by using the avidin–biotin complex method (Vector Laboratories), including heat-induced antigenretrieval procedures. Incubation with antibodies against CK5 (1:150; #71536, Cell Signaling), TIMP1 (1:150; abs149999, Absin), EZH2 (1:100; #5246, Cell Signaling), SF3B2 (1:150; abs116920, Absin), COPS8 (1:150; abs143431, Absin) was carried out at 4 °C for 12 h. PanCK CK5 was used to mark cancer cells. DAPI was used to highlight all nuclei. Visualization and quantitation of the different fluorophoreswas achieved on the TissueFAXS Spectra Systems and StrataQuest analysis software (TissueGnostics); the location information and expression of all the markers were computed for every cell were recorded.

Statistical analysis

Statistical analyses were performed using python scipy packages with unpaired Student's t-test or Wilcoxon test. All p-values were adjusted for multiple testing using the Bonferroni correction p values lower than 0.05 are considered statistically significant. ∗, ∗∗, ∗∗∗ and ∗∗∗∗ indicate p < 0.05, p < 0.01, p < 0.001 and p < 0.0001, respectively.

Ethics

The study protocol was approved by the Medical Ethics Committee of Tongji Medicine College of HUST (ID: S1025). Written informed consents were obtained from all participants.

Role of funders

All the funders had important roles in study design, data collection, data analysis and writing of report.

Results

Single-cell transcriptome atlas of normal cervix, precancerous cervical lesions, and cervical cancer

To characterize the multicellular ecosystem of cervical premalignant and malignant lesions, scRNA-seq was conducted on 20 biopsies, including five healthy controls (CON), four HSIL, five MIC, and six ICC samples (Fig. 1a–e; Fig. S1; Table S1). The patients did not receive chemotherapy or radiotherapy prior to biopsy. In total, ∼122,400 cells were retained for analysis after quality control (Fig. S2; Table S2). Among these cells, 33,622 cells (27.46%) were from the CON group, 20,880 cells (17.07%) were from the HSIL group, 29,068 cells (23.74%) were from the MIC group, and 38,848 cells (31.73%) were from the ICC group. Using graph-based uniform manifold approximation and projection (UMAP), we identified 29 high-confidence cell clusters (Fig. 1e). Based on canonical gene marker expression (Fig. 1f and g), we identified nine major cell types: T cells, B cells, squamous epithelial cells, columnar epithelial cells, macrophages, neutrophils, mast cells, fibroblasts, and endothelial cells (Fig. 1b). The relative cell lineage abundance showed distinct stepwise changes from the CON group to the ICC group (Fig. 1h and i). The relative abundances of T cells, B cells, macrophages, neutrophils, and mast cells were stable across the four groups. The proportion of columnar epithelial cells, fibroblasts, and endothelial cells decreased from the CON group to the ICC group. Additionally, there was a gradual increase in the percentage of squamous epithelial-like cells. These results suggest that the multicellular cervical ecosystem exhibits progressive changes during the transition from normal cervix to precancerous lesions and then to cervical cancer.

Fig. 1.

Single-cell atlas of normal cervix, precancerous cervical lesions, and cervical cancer. (a) Schematic showing scRNA-seq experimental design. In total, ∼122,400 cells from 20 cervical biopsies (5 healthy control [CON], 4 high-grade intraepithelial neoplasia [HSIL], 5 microinvasive carcinoma of the cervix [MIC], and 6 invasive cervical squamous carcinoma [ICC]) were included in the study. (b) UMAP plot of 9 major cellular populations in ∼122,400 cells identified by the expression of known marker genes, including T cells, B cells, squamous epithelial cells, columnar epithelium cells, macrophages, neutrophils, mast cells, fibroblasts, and endothelial cells. Each dot represents single cell and is colored according to cell type. The number of cells is marked in parentheses. (c) UMAP plots of cells clustered by sample origin. (d) UMAP plots of cells clustered by group origin. (e) UMAP plot of 28 cellular clusters identified in ∼122,400 cells. (f) Canonical cell markers were used to label clusters with cell identity, as represented in the UMAP plot. (g) Violin plots of representative genes used to identify major cell types. (h) Average proportion of 9 major cell types among CON, HSIL, MIC, and ICC samples. (i) Proportion of 9 major cell types in each sample among CON, HSIL, MIC, and ICC groups.

Identification and functional characterization of malignant cells

Epithelial-like cells were further divided into 16 clusters (Fig. 2a–e). We observed a trend of proportional changes in epithelial-like cell clusters across the four groups that revealed four characteristics (Fig. 2c; Tables S2–S5): (1) Clusters C0, C4, C6, C10, and C25 were almost exclusively present in the HSIL, MIC, and ICC groups. The epithelial-like cell fractions gradually increased with lesion progression and accounted for most epithelial-like cell clusters in the MIC and ICC groups. Thus, we initially considered these cell clusters to be tumor cells. (2) Clusters C3, C7, C12, C14, C24, C26, and C28 comprised most of the epithelial-like cell clusters from the CON and HSIL groups. However, the proportion of these cells among epithelioid cells from the MIC and ICC groups was extremely low. Therefore, these cells were considered to be potential normal epithelial cells. (3) Clusters C19, C16, and C15 were epithelial cells that were unique to the HSIL group. (4) The proportion of cluster C2 was relatively stable among epithelial-like cell clusters of the four groups.

Fig. 2.

Identification and functional characterization of malignant cells. (a) UMAP plots of epithelial-like cells colored by malignant/non-malignant cells (top) or by cluster (bottom). (b) UMAP plots showing the expression and distribution of canonical cervical cancer markers among epithelial-like cell clusters. (c) Heatmap (left) and line graphs (right) showing the proportion changes in epithelial-like cell and malignant cell clusters across the CON, HSIL, MIC, and ICC groups. Clusters C0, C4, C6, C10, and C25 were exclusively detected in the HSIL, MIC, and ICC groups. Their fractions gradually increased with lesion progression and accounted for most epithelial-like cell clusters in the MIC and ICC groups. Clusters C3, C7, C12, C14, C24, C26, and C28 comprised most of the epithelial-like cell clusters from the CON and HSIL groups, while their proportions in epithelioid cells from the MIC and ICC groups were extremely low. Clusters C19, C16, and C15 were epithelial cells that were unique to the HSIL group. The proportion of cluster C2 was relatively stable among epithelial-like cell clusters from the 4 groups. (d) Dot plot showing average canonical cervical cancer marker expression across different epithelial-like cell and malignant cell clusters. Cervical cancer markers were highly expressed in clusters C0, C2, C4, C6, C10, and C25. The dot size is proportional to the fraction of marker-expressing cells in each group. (e) Heatmap showing differences in CancerSEA signatures scored by ssGSEA per cell among malignant cells (C0, C4, C6, C10, and C25), potential C2 cervical stem cells, and HSIL-specific epithelial-like cells (C15, C16 and C19). DNA repair, cell cycle, DNA damage, invasion, and proliferation signatures strongly appeared in cluster C2. Clusters C0, C4, and C6 showed enriched angiogenesis, inflammation, hypoxia, differentiation, EMT, and metastasis. The HSIL-specific clusters C15 and C16 exhibited a transition state between C2 and tumor cell clusters C0, C4, C6, and C10. (f) Heatmap showing large-scale CNVs in single cells, inferred from scRNA-seq data. Non-malignant cell cluster C3 was treated as the reference (top). Large-scale CNVs were observed in malignant cells (bottom). CNV analysis demonstrated positive CNV scores for C0, C4, C6, C10, and C25. CNV scores of C2 from the CON and HSIL groups were negative, whereas those from the MIC and ICC groups were positive. (g) Violin plot showing CNV scores distributions in C2 sub-clusters by group. C2 CNV scores from the CON and HSIL groups were negative, whereas those from the MIC and ICC groups were positive. Non-malignant cell cluster C3 was used as the reference. (h) Top 12 upregulated hallmark pathways identified via ssGSEA in malignant cell clusters compared to other epithelial-like cells. The red bar indicates that the q value was less than 0.05.

To further characterize the detected tumor cells, we examined cervical cancer-specific marker expression in each epithelial cell cluster. Most of the cervical cancer markers were highly expressed in clusters C0, C4, C6, C10, and C25, further indicating that these clusters represent tumor cells (Fig. 2b–d; Fig. S3a). Notably, cluster C2 also highly expressed large amounts of cervical cancer markers. This expression was higher than in other tumor cells (Fig. 2b–d; Fig. S3a).

The CNV analysis demonstrated positive CNV scores for C0, C4, C6, C10, and C25, thus corroborating the above hypothesis that these cell clusters were tumor cells (Fig. 2f). However, the CNV scores of C2 in the CON and HSIL groups were negative, while the CNV scores of C2 from the MIC and ICC groups were positive (Fig. 2f and g). Moreover, the CNV score distribution of malignant cells in our study and the CNV distribution of cervical cancer bulk samples in the cancer genome atlas (TCGA) database was similar (Fig. S3b). As C2 was the only cell cluster with a relatively stable epithelial cell proportion (6–15%) among the four stages (Table S3) and had a unique CNV score distribution (Fig. 2f and g), we concluded that C2 cells were putative cervical CSCs.

Next, we performed functional analysis of the identified malignant cell clusters (C0, C4, C6, C10, and C25) and the potential CSC cluster (C2) (Fig. 2e–h; Fig. S3c and S3d; Table S6). Single-sample gene set enrichment analysis (ssGSEA) revealed that E2F targets, MYC targets, G2M checkpoint, mitotic spindle, glycolysis, and spermatogenesis pathways were significantly enriched in cluster C2 compared to other epithelial-like cell clusters (Fig. 2h; Fig. S3d).^11,12 The malignant cell clusters C0, C4, C6, C10, and C25 showed enriched hallmark pathways, including interferon pathways [interferon (IFN)-α and IFN-γ responses], MYC targets, and epithelial–mesenchymal transition (EMT) (Fig. 2h; Fig. S3c; Table S6). Enriched complement system, IL6 JAK-STAT3 signaling, and inflammatory response markers in cluster C10 suggested their potential involvement in immune responses during cervical cancer progression (Fig. 2h; Fig. S3c; Table S6). The KRAS pathway plays an influential role in tumor evolution¹³ and was enriched in clusters C0, C6, and C10 (Fig. 2h; Fig. S3c; Table S6). In the CancerSEA signature enrichment analysis, DNA repair, cell cycle, DNA damage, invasion, and proliferation signatures were most strongly enriched in cluster C2 (Fig. 2e).¹⁴ Malignant cell clusters C0, C4, and C6 were enriched in angiogenesis, inflammation, hypoxia, differentiation, EMT, and metastasis (Fig. 2e). HSIL-specific clusters C15 and C16 exhibited a transitional state that shared the characteristics of C2 and tumor cell clusters C0, C4, C6, and C10 (Fig. 2e).

Identification and characterization of cervical cancer stem cells

To investigate the stem cell properties of cluster C2, we calculated the mRNA expression-based stemness index (mRNAsi) for epithelial and malignant cell clusters. This analysis revealed that C2 had the highest mRNAsi scores (Fig. 3a; Table S7).⁹ Notably, mRNAsi scores analysis showed a gradual decrease in the mRNAsi score of C2 with lesion progression (Fig. 3b). Moreover, cluster C2 exhibited high expression of the core stem cell factor enhancer of zeste 2 polycomb repressive complex 2 subunit EZH2 (Fig. 2b–d).15, 16, 17, 18 We also compared enriched stem cell-related gene signatures across epithelial and tumor cell clusters and observed significantly enriched gene signatures in C2 (Fig. S3c and S4a).^9,19, 20, 21 Furthermore, trajectory analysis revealed that C2 was the starting point for the evolution of tumor cell clusters (Fig. 3d). In summary, C2 was the only cluster with a relatively stable proportion of epithelial-like clusters (6–15%) across all stages (Fig. 2c), characteristic CNV scores (negative for C2 from the CON and HSIL groups but positive for C2 from the MIC and ICC groups) (Fig. 2f and g), cervical cancer marker expression profiles (Fig. 2b–d), and stem cell-related gene signature enrichment (Figs. S3c and S4a). We suggest that cluster C2 with high EZH2 expression might be cervical CSCs.

Fig. 3.

Identification and characterization of cervical cancer stem cells. (a) Violin plot showing mRNA expression-based stemness index (mRNAsi) distributions for epithelial and malignant cell clusters (rank-sum test). C2 had the highest mRNAsi scores. The HSIL group-specific epithelial clusters C15 and C19 had higher mRNAsi scores than most of the epithelial clusters. The mRNAsi scores of C16 exhibited a bipolar distribution, with the higher scoring fraction having mRNAsi scores close to those of C2. (b) Violin plot showing mRNAsi distributions in cluster C2 according to the groups, which revealed that the C2 mRNAsi score gradually decreased with lesion progression (two sample t-test). (c) Heatmap showing the enriched stem cell-related gene signatures across epithelial and tumor cell clusters. Cluster C2 showed significant enrichment of stem cell-related gene signatures. (d) Developmental trajectory of malignant cells (C0, C4, C6, C10, and C25), cervical stem cells (C2), and HSIL-specific epithelial-like cells (C15, C16, and C19). C2 was the starting point for the evolution of tumor cell clusters. The HSIL group-specific epithelial clusters C15, C16, and C19 could possibly be transient-amplifying cells (TAC), which acted as a transition state between cancer stem cells and cancer stem cell-derived tumor cells. (e) Heatmap showing enriched CancerSEA signatures in cluster C2 according to groups identified via ssGSEA analysis. Proliferation, cell cycle, DNA damage, and DNA repair signatures were significantly enriched in cluster C2 of the CON group. Cluster C2 in the ICC group was significantly enriched in invasion, metastasis, EMT, angiogenesis, and hypoxia signatures, whereas that in the HSIL and MIC groups showed a transitional state between the CON and ICC groups. (f) Heatmap showing the enriched hallmark pathways of cluster C2 according to groups identified via ssGSEA analysis. Cluster C2 of the CON group showed significant enrichment of G2M checkpoint, DNA repair, MYC targets, E2F targets, and oxidative phosphorylation pathways. C2 of the ICC group showed significant enrichment of EMT, angiogenesis, IL-6, Wnt, TGF-β, glycolysis, and hypoxia pathways. C2 in the HSIL and MIC groups showed a transitional state between the CON and ICC groups. (g) Bar graph showing the cell cycle distribution for cluster C2 by group. The proportion of G1 phase cells in C2 gradually increased as the lesion progressed. (h) Kaplan–Meier plot showing that CESC patients (n = 292) from the TCGA dataset with high cluster C2 marker (C2 top 30 marker gene signature) expression had significantly lower progression free survival (p = 0.012) (log-rank test). The high and low C2 marker expression groups are divided by the 50% quantile value of the mean expression of the above gene set (cut-off value = 50%).

We next explored the functional changes in C2 stem cells during malignant transformation from normal cervix tissue to invasive cervical carcinoma. The functional changes in cervical CSCs were characterized by two major features: the original non-malignant stem cell properties characterized by high proliferation and low differentiation gradually diminished, and CSC properties, characterized by EMT, invasion, and metastasis, progressively enhanced (Fig. 3e and f). Specifically, the proliferative properties of C2 gradually declined during transformation. ssGSEA analysis of the CancerSEA signature, proliferation, cell cycle, DNA damage, and DNA repair signatures were significantly enriched in cluster C2 of the CON group (Fig. 3e). In addition, ssGSEA of hallmark signatures had similar results, demonstrating significant enrichment of G2M checkpoint, DNA repair, MYC targets, and E2F target pathways in cluster C2 of the CON group (Fig. 3f). Meanwhile, the expression levels of six genes involved in cell cycle and cell proliferation (RRM2, TK1, BIRC5, MKI67, TOP2A, and ZWINT) in cluster C2 gradually decreased as lesions progressed (Fig. S4b).22, 23, 24, 25, 26, 27 Similar findings were observed in the cell cycle analysis, where the proportion of G1 phase markers in C2 gradually increased as the lesion progressed, suggesting a diminished proliferative capacity and enhanced C2 differentiation (Fig. 3g; Table S8).

In contrast, cluster C2 of the ICC group was significantly enriched for invasion, metastasis, EMT, and angiogenesis signatures, suggesting enhanced invasive and metastatic properties of C2 in the ICC group (Fig. 3e). This result was verified via hallmark gene set enrichment analysis, in which EMT and angiogenesis were significantly enriched in cluster C2 in the ICC group (Fig. 3f). In addition, ssGSEA of hallmark signatures revealed high expression of EMT-promoting pathways, including IL-6, Wnt, and TGF-β in the ICC group C2 (Fig. 3f).^28,29 Meanwhile, the enriched energy metabolism signatures suggested an enhanced glycolysis signature in C2 of the ICC group and a stronger oxidative phosphorylation signature in C2 of the CON group (Fig. 3f). Highly proliferative cells have a greater OXPHOS capacity, whereas EMT promotes glycolysis, which subsequently facilitates tumor cell metastasis.³⁰ EMT score analysis also indicated a progressive EMT score enhancement in C2 from the CON to the ICC group (Fig. S4c).¹⁰ Comparing the mRNAsi scores among cluster C2 from the four groups demonstrated a gradual decrease in the mRNAsi score of C2 with lesion progression (Fig. 3b). Notably, the stemness signature of CancerSEA was significantly enriched in cluster C2 in the ICC group (Fig. 3e). Thus, we concluded that the mRNAsi score better indicated non-malignant stem cell properties, whereas the stemness signature of CancerSEA better evaluated tumor stem cells. In addition, differentiation, quiescence, and hypoxia signatures were significantly enriched in cluster C2 of the ICC group (Fig. 3e and f). Hypoxia induces EMT in tumors and promotes tumor cell invasion and metastasis,²⁸ whereas quiescence is associated with tumor cell drug resistance.³¹ C2 in the HSIL and MIC groups showed a transition state between the CON and ICC groups (Fig. 3e and f). In particular, the functions of C2 in the CON and HSIL groups were similar, while C2 in the MIC and ICC groups shared common features (Fig. 3e and f). Survival analysis using data from the TCGA database revealed that cervical cancer patients with high expression of C2 marker gene signatures had significantly lower progression free survival than those with low expression of C2 signatures (p = 0.012) (log-rank test), suggesting that cervical cancer patients with high expression of C2 markers are more vulnerable to recurrence (Fig. 3h; Table S9).

The HSIL group-specific epithelial clusters C15, C16, and C19 could be transient-amplifying cells (TAC) that act as a transition state between CSCs and CSC-derived tumor cells.^30,32 This finding was confirmed via trajectory analysis (Fig. 3d) and ssGSEA of the tumor-related signatures of malignant cell clusters (Fig. S2e and S3c). Cervical cancer marker genes were highly expressed in C15, C16, and C19 (Fig. 2d). Notably, comparison of mRNAsi scores among epithelial clusters revealed that C15 and C19 had higher scores than most of the epithelial clusters, whereas the mRNAsi scores of C16 exhibited a bipolar distribution. The higher scoring fraction had mRNAsi scores close to C2 (Fig. 3d; Table S7). These results strongly suggest an association between C16 cells and C2 stem cells. Previous studies reported TAC as a highly proliferative cell type.³² This finding was also reflected in our study. ssGSEA of CancerSEA signatures revealed enriched proliferation, cell cycle, DNA damage, and DNA repair signatures in C15 and C16 (Fig. 2e). In addition, a high percentage of C15 and C16 in the G2M and S phases were consistent with high proliferation characteristics (Fig. S4d; Table S10).

Our results provide a comprehensive picture of malignant C2 cervical stem cell transformation, thereby providing a valuable paradigm describing non-malignant stem cell transformation into CSCs. The CNV score of C2 in the HSIL group was negative, whereas that of cluster C2 in the MIC group was positive, suggesting that malignant transformation of C2 cervical stem cells occurs between the HSIL and MIC stages (Fig. 2f and g). This result is consistent with the current understanding that the long-term persistence of high-risk HPV leads to genomic instability in cervical stem cells. This instability causes somatic mutations and eventual malignant transformation.³³

Since the above conclusions were based on bioinformatic analysis, further validation was needed to provide sufficient proof for the existence of stem-like C2 cluster cells. Thus, we validated these results in a cervical cancer tissue microarray (TMA) cohort. A total of 85 cervical cancer samples with 22 matched cases of paracancerous tissues were included for subsequent analysis based on the results of nuclear immunohistochemical staining of four top markers of C2 cluster (TIMP1, EZH2, SF3B2, COPS8) (Table S9). Our findings confirm the existence of C2 cluster cells (Fig. 4a). Also, survival analysis showed that a high percentage of C2 cluster cells was associated with worse overall survival (Fig. 4b) and disease-free survival (Fig. 4c), which is consistent with the results of the survival analysis of TCGA data (Fig. 3h). The proportion of C2 cluster cells in the cancer lesions was significantly higher than that of paired paracancerous tissues (Fig. 4d). Patients with recurrence and those with lymphatic metastases also had a higher proportion of C2 cluster cells in tumor lesions than those without recurrence and those without lymph node metastases, respectively, although there was no statistical difference (Fig. 4e and f). There was no difference in the C2 cluster cell rate in cancer lesions of different pathological subtypes (Fig. 4g). Taken together, these results confirm the existence of stem-like C2 cluster cells and indicate that C2 cluster correlates with neoplastic recurrence.

Fig. 4.

Verification of high C2 cluster cell percent correlates with poor patient prognosis. (a) Representative images for multiplexed immunofluorescence staining of C2 cluster cells in cervical cancerous tissues (CK5, green; TIMP1, cyan; EZH2, yellow; SF3B2, red; COPS8, violet). PanCK CK5 was used to mark cancer cells. DAPI was used to highlight all nuclei. Scale bars = 0.01 mm. (b) Overall survival analysis for cervical cancer patients (n = 85) with different proportions of C2 cluster cells in cancerous tissues. Differences between groups were evaluated using log rank test (p = 0.03) (cut-off value = 50%). (c) Disease free survival analysis for cervical cancer patients (n = 85) with different proportions of C2 cluster cells in cancerous tissues. Differences between groups were evaluated using log rank test (p = 0.03) (cut-off value = 50%). (d) Comparison of C2 cluster cell frequency between cancerous and matched paracancerous tissues from cervical cancer patients (p = 0.046) (rank-sum test). The proportion of C2 cluster cells was calculated using the ratio of cells co-expressing C2 maker genes (TIMP1, EZH2, SF3B2, COPS8, CK5) to the number of overall cells in the samples. (e) Comparison of C2 cluster cell frequency in cancerous tissues between relapsing patients and non-relapsing patients (p = 0.0806) (rank-sum test). (f) Comparison of C2 cluster cell frequency in cancerous tissues between patients with lymph node metastasis (LNM) and those with no lymph node metastasis (Non-LNM) (p = 0.0754) (rank-sum test). (g) Comparison of C2 cluster cell frequency in cancerous tissues between patients with squamous carcinoma and those with other pathological types (p = 0.955) (rank-sum test).

Deciphering the functional heterogeneity of malignant cells in HSIL, MIC, and ICC

We found that the proportion of tumor cell clusters C0, C4, C6, and C10 gradually increased with lesion progression (Fig. 2c). Thus, we investigated functional changes in malignant cells across the HSIL, MIC, and ICC stages. CNV analysis was conducted on only a random selection of cells. Because the malignant cell clusters from HSIL constituted only a small percentage of the total tumor cells, these cells might not be selected (Table S3). Therefore, we performed CNV analysis on HSIL-derived tumor cells and found that these cells exhibited abnormal CNV signals, confirming tumor cell identity (Fig. 5a and b). This result indicates that a few tumor cells are present in the HSIL stage, although these cells constitute a very low percentage of epithelial-like cell clusters (Tables S2 and S3).

Fig. 5.

Characterization of the functional heterogeneity of malignant cells. (a) Heatmap and (b) violin plot showing the distributions of CNV scores in HSIL-derived tumor cells (C0, C4, and C10). (c) Heatmap depicting pairwise correlations of intratumoral programs derived from HSIL (left), MIC (middle), and ICC (right). Nine prominent meta-programs were identified when these samples were considered together, including cell cycle, epithelial–mesenchymal transition (EMT), epithelial differentiation (Epi dif. 1 and Epi dif. 2), LncRNA (LncRNA-rich 1 and LncRNA-rich 2), stress responses (stress), IFN-response genes (IFN-response), and hypoxia signatures. Only two meta-programs were identified in the HSIL group alone (cell cycle and Epi dif. 1). Five meta-programs were identified in the MIC group (cell cycle, EMT, LncRNA-rich 1, Epi dif. 1, Epi dif. 2, and IFN-response). All nine meta-programs were observed in the ICC group. These observations demonstrate that malignant cell heterogeneity increases along with disease progression. (d) Kaplan–Meier plot showing that CESC patients from the TCGA dataset with high hypoxia signature expression had significantly lower progression free survival (log-rank test). The high and low groups are divided by the 50% quantile value of the mean expression of the above gene set. (e) Enriched CancerSEA signatures of malignant cell clusters (C0, C4, C6, and C10) across HSIL, MIC, and ICC groups. (f) Heatmap showing enriched hallmarks pathways in malignant cell clusters (C0, C4, C6, and C10) across HSIL, MIC, and ICC groups.

Non-negative matrix factorization was applied to identify potential gene co-expression programs in malignant cells from HSIL, MIC, and ICC samples.³⁴ Nine prominent meta-programs were identified when these samples were considered together: cell cycle, EMT, epithelial differentiation (Epi dif. 1 and Epi dif. 2), lncRNA (lncRNA-rich 1 and lncRNA-rich 2), stress response (stress), IFN-response, and hypoxia (Fig. S5a). However, in the HSIL group alone, only two meta-programs were identified, including cell cycle and Epi dif. 1 (Fig. 5c). Six meta-programs were identified in MIC, which included cell cycle, EMT, lncRNA-rich 1, Epi dif. 1, Epi dif. 2, and IFN-response (Fig. 5a). In the ICC group, all nine meta-programs were observed (Fig. 5a). These observations demonstrate that malignant cell heterogeneity increases with disease progression. In particular, the overall survival of cervical cancer patients with high expression of hypoxia meta-signature genes was significantly lower than that of patients with low hypoxia-related expression in the TCGA dataset (Fig. 5d). Hypoxia is closely associated with tumor proliferation, invasion, metastasis, and angiogenesis.^28,35

In addition, comparing the malignant cell clusters from the HSIL, MIC, and ICC samples via ssGSEA of tumor-related signatures revealed that EMT, metastasis, angiogenesis, quiescence, inflammation, hypoxia, and differentiation in malignant cells gradually increased concomitant with lesion progression (Fig. 5e and f; Fig. S5b). Further characterization of the metabolic profiles suggested that malignant cells in the HSIL, MIC, and ICC groups had distinct metabolic patterns (Fig. S5c).³⁶ Overall, our results demonstrate increased tumor cell heterogeneity and malignancy with lesion progression.

Characterization of high-risk HPV gene expression in malignant cells and cancer stem cells

We detected the expression of high-risk HPV (HR-HPV) transcripts in all epithelial and malignant cell clusters (Fig. 6a and b; Fig. S6a). As expected, no cells from the CON group expressed HR-HPV genes (Fig. 6c; Table S11). In general, the proportions of epithelial cells and malignant cells from the HSIL, MIC, and ICC groups that expressed at least one HR-HPV transcript (irrespective of strain) were 23.05%, 39.23%, and 43.69%, respectively. In addition, we identified distinct expression characteristics of HR-HPV transcripts in C2 CSC and malignant cell clusters at different disease stages, along with common features (Fig. 6d). C2 CSC and malignant cell clusters (C0, C4, C6, and C10) from the HSIL group highly expressed E5, whereas high E7 expression was observed in these clusters in the MIC stage. In addition, C2 CSCs and most tumor cell clusters (including C0, C4, and C6) in the ICC group exhibited high E6 expression. Similar results were obtained via integration analysis of malignant cell clusters (Fig. 6d).

Fig. 6.

Gene expression profiles of high-risk HPV genes and HPV gene-related host genes. (a) UMAP plots showing the expression and distribution of HPV transcripts among epithelial cell and malignant clusters. (b) UMAP plots showing E5, E6, and E7 expression and distribution among epithelial cell and malignant clusters. (c) Heatmap depicting the HPV transcript-positive cell percentage among epithelial cells, cancer stem cells, and malignant clusters. (d) Heatmap demonstrating HPV transcript expression among cancer stem cell and malignant clusters. Cancer stem cells (C2) and malignant cell clusters (C0, C4, C6, and C10) from the HSIL group highly expressed E5; however, high E7 expression was observed in these clusters from the MIC stage. Moreover, cancer stem cells (C2) and most tumor cell clusters (including C0, C4, and C6) in the ICC group showed high E6 expression. (e) Heatmap showing the correlation between HPV transcript expression and host gene expression in malignant cell and cancer stem cell clusters. (f) Kaplan–Meier plot showing that CESC patients (n = 292) from the TCGA dataset with high expression of E5-related (upper), E6-related (middle), and E7-related (bottom) host gene sets had significantly higher overall survival (log-rank test). The high and low groups are divided by the 50% quantile value of the mean expression of the above gene sets.

Our study revealed associations between HR-HPV transcript expression and host gene expression in CSCs and tumor cell clusters (Fig. 6e) that differed among the HSIL, MIC, and ICC stages (Fig. S6b). Overall, E5-related host genes were enriched in C2 CSC and malignant cell clusters from the HSIL group, while E6 and E7-related host genes were enriched in the same clusters in the ICC group (Fig. S6b). MIC group-derived CSC and malignant cell clusters showed transitional features. Notably, functional analysis of E5-, E6-, and E7-related host genes revealed associations with immune function (Fig. S6c). Survival analysis revealed that cervical cancer patients with high expression of these gene sets had significantly higher survival rates than patients with low expression (Fig. 6f), indicating that HPV-related host genes might be related to immunity against HPV and tumors.

T cell clustering and functional analysis in CON, HSIL, MIC, and ICC

We analyzed gene expression in 18,697 T cells categorized into 10 clusters (Fig. 7a and Fig. S7a). We also investigated classical T cell marker expression and distribution to annotate the clusters (Fig. 7b–c and Fig. S7b). The distribution of T cells derived from the four groups exhibited continuity (Fig. 7a and Fig. S7a). Cluster T4 was identified as regulatory T cells (Tregs) via high FOXP3, IL2RA, and IKZF2 expression (Fig. 7b–c and Fig. S7b). Meanwhile, cluster T6 had high ICOS, PPP2R5C, and CD4 expression indicating T helper cell (Th) identity (Fig. 7b–c and Fig. S7b). The cluster Treg-T4 proportion tended to increase among the four groups, whereas the cluster Th-T6 proportion gradually decreased from the CON group to the ICC group. These observations indicate that the immune microenvironment among the four groups gradually became biased toward immunosuppression.

Fig. 7.

Functional states of cervix-infiltrating T cells in the CON, HSIL, MIC, and ICC groups. (a) UMAP projection of 18,697 T cells demonstrating 10 main clusters (upper) and group distributions (bottom). (b) UMAP plots showing classical T cell marker expression and distribution among T cell clusters. (c) Expression of T cell-specific markers across different clusters. The dot size is proportional to the fraction of marker-expressing cells in the group. (d) Line graph showing the relative T cell abundance across the CON, HSIL, MIC, and ICC groups. The proportion of cytotoxic effector T cell cluster CD8-T0 was significantly low in the ICC group. The activation-coupled exhausted CD8+ T cell clusters T2 and T3 were significantly increased in the MIC and ICC groups. (e) Heatmap showing enriched cytotoxic and exhaustion signatures of T cell clusters. CD8+ T cells clusters (T2 and T3) exhibited an activation-coupled exhaustion program. (f) Kaplan–Meier plot showing that CESC patients from the TCGA dataset with high CD8-T2 signature (upper) and CD8-T3 signature (bottom) expression had significantly lower progression free survival (log-rank test). The high and low groups are divided by the 75% quantile value of the mean expression of the above gene sets. (g) Developmental trajectory of CD8+ T cell clusters (T0, T2, T3, T7, and T9) revealed a binary branched structure with a connection between two end states: naïve T cells CD8-T9 were the starting point, CD8-T0/T7 cells were the end state of cytotoxic T cells, and CD8-T2/T3 cells were the endpoint of activation-coupled exhausted T cells. Specifically, there might be a connection between the 2 end states, in which cytotoxic T cells can transform into activation-coupled exhausted state T cells.

T cells from T0, T2, T3, T7, and T9 exhibited high CD8A and CD8B expression indicating these cells are CD8+ T cells, which were the main T cell clusters in this study (Fig. 7b–d and Fig. S7b). For CD8+ T cell clusters, cluster CD8-T9 highly expressed LEF1 and TCF7 and were identified as naïve CD8+ T cells (Fig. 7b–c and Fig. S7b). Clusters CD8-T0 and CD8-T7 corresponded to effector T cells owing to high expression of cytotoxic markers, such as NKG7 and GZMA (Fig. 7b–c and Fig. S7b). Notably, clusters CD8-T2 and CD8-T3 exhibited high cytotoxic marker (NKG7, GZMA, GZMH, and KLRD1) and exhaustion marker expression (TIGIT, HAVCR2, and LAG3), indicating the presence of an activation-coupled exhaustion program (Fig. 7b–c and Fig. S7b). This result was consistent with enrichment analysis using cytotoxic and exhaustion signatures^37,38 (Fig. 7e). A similar phenomenon was reported in EBV-associated nasopharyngeal carcinoma,³⁹ in which tumor-infiltrating T cells are highly activated and exhausted. This specific phenomenon might be related to chronic HR-HPV infection and could be a common feature of virus-associated tumors. Survival analysis revealed that TCGA cervical squamous carcinoma (CESC) patients with high CD8-T2 and CD8-T3 signature gene expression had significantly lower progression-free survival than patients with low expression (Fig. 7f). Regarding the CD8+ T cell subtype composition among the four groups, the proportion of the cytotoxic effector T-cell cluster CD8-T0 was significantly lower in the ICC group. However, for activation-coupled exhausted CD8+ T cell clusters, CD8-T2 was significantly increased in the MIC and ICC groups. CD8-T3 was primarily identified in the ICC group (Fig. 7d). These results indicate that activation-coupled exhausted CD8+ T cells play a key role in establishing the immune microenvironment in cervical cancer, which emerged at the MIC stage. Trajectory analysis of CD8+ T cells revealed a binary branched structure with a connection between two end states. Naïve T cells CD8-T9 formed the starting point, CD8-T0/T7 were the end-state cytotoxic T cells, and CD8-T2/T3 were the endpoints of activation-coupled exhausted T cells. We observed a potential connection between the two end states, in which cytotoxic T cells can transform into activation-coupled exhausted T cells (Fig. 7g). This result is consistent with those of previous studies of the tumor immune microenvironment.^40,41

Cervical microenvironment composition and cell–cell interactions in CON, HSIL, MIC, and ICC

To characterize the intercellular interactions in the cervical microenvironment, we investigated cell–cell interactions based on ligand–receptor gene expression in single cells (Fig. 8a and Fig. S8a). We identified several angiogenesis-related interactions between CSC and endothelial cells, including VEGFA-FLT1, VEGFA-KDR, NRP2-SEMA3F, and NOTCH2-JAG2 (Fig. 8b and Fig. S8d). Interactions between cervical stem cells and fibroblasts included EMT (WNT2-FZD1, WNT5A-ANTXR1, WNT5A-FZD5, and WNT5A-PTPRK), proliferation, and metastasis promotion (Fig. 8c and Fig. S8d). Cervical stem cells induced M2 macrophage activation via INHBA, EGFR-AREG, LIFR-OSM, MDK-LRP1, MDK-SORL1, OSMR-OSM, TNFSF10-TNFRSF10D, and LGALS9 interactions (Fig. 8d and Fig. S8d). Next, we identified immune–inhibitory interactions between cervical stem cells and T cells, including HLA-E-KLRC1 and LGALS9-HAVCR2 (Fig. 8e and Fig. S8d). Both the number and intensity of these interactions increased with cervical lesion progression (Fig. 8c and Fig. S8d).

Fig. 8.

Composition and intercellular interactions of the cervical microecosystem in the CON, HSIL, MIC, and ICC groups. (a) Circles plots showing putative intercellular interactions among different cell types in the CON, HSIL, MIC, and ICC groups. The directional strings represent cell–cell interactions based on ligand expression in one cell type and corresponding receptor expression in another cell type. The width of each string is proportional to the amount of different interaction pairs. Each color corresponds to a cell type. (b–e) Dot plot demonstrating the mean expression level of selected ligand–receptor interactions between CSC and endothelial cells (b), fibroblasts (c), macrophages (d), and T cells (e). The color gradient represents the interaction level. The size of each circle indicates the P value (permutation test). (f) Kaplan–Meier plot (log-rank test) showing that the overall survival was similar between the G1 group (HSIL-like) and G2 group (MIC-like). Both groups showed significantly higher overall survival than G3 group (ICC-like) in CESC patients from the TCGA dataset. (g) Box plots showing the IFN-α (upper) and IFN-γ (bottom) scores of individual cells in cellular cervical microenvironment compositions across the CON, HSIL, MIC, and ICC groups. The results revealed significantly high IFN scores in the ICC group for most major cell types, except endothelial cells (rank-sum test).

Based on the cell type-specific markers obtained from scRNA-seq analysis of the cervical microenvironment, we analyzed TCGA CESC data by characterizing the marker expression patterns via bulk RNA-seq. We analyzed 292 untreated CESC patients and divided them into three groups according to the differences in the cervical microenvironment composition: G1 (HSIL-like), G2 (MIC-like), and G3 (ICC-like) (Fig. S8b–c). Survival analysis revealed that the overall survival was similar between the G1 group (HSIL-like) and G2 group (MIC-like). G1 and G2 survival were significantly higher than G3 survival (ICC-like) (Fig. 8f). This result suggests that the MIC cervical microenvironment might be more similar to HSIL than ICC, when the cervical microenvironment is considered as a whole.

Notably, we calculated and compared IFN scores (IFN-α and IFN-γ scores) for each major cell type among the four groups. The results revealed significantly higher IFN scores in the ICC group for most major cell types, except the endothelium, suggesting a globally upregulated IFN responses in the cervical cancer microecosystem (Fig. 8g). This feature was previously reported in EBV-associated nasopharyngeal carcinoma,³⁹ which further supports the conclusion that viral infection plays an important role in shaping the virus-associated tumor microenvironment.

Discussion

In this study, we comprehensively investigated malignant and immune heterogeneity, dynamic variations in cellular composition, and cell–cell interactions in the cervical ecosystem across CON, HSIL, MIC, and ICC (Fig. 9). Notably, our data comprehensively describe the malignant transformation process of cervical CSC, thus offering a valuable paradigm for understanding how adult stem cells become malignant tumor stem cells. Our data demonstrate that the malignant transformation of cervical stem cells occurs between the HSIL and MIC stages.

Fig. 9.

Schematic of molecular and phenotypic signature evolution during preneoplastic cervical cancer progression. The functional changes in cervical cancer stem cells during malignant transformation are characterized by two major features: the original non-malignant stem cell properties characterized by high proliferation and low differentiation gradually diminish, and the cancer stem cell properties characterized by EMT, invasion, and metastasis progressively increase. Tumor cell heterogeneity and malignancy increase with lesion progression. For T cells, the proportion of Tregs increases, whereas that of Th gradually decreases with lesion progression. As the disease progresses, the proportion of cytotoxic effector CD8+ T cells decreases, whereas the relative abundance of activation-coupled exhausted CD8+ T cells increases. High IFN scores were observed in the ICC group for most major cell types, except the endothelium, suggesting globally upregulated IFN responses in the cervical cancer microecosystem.

CSCs, also known as tumor-initiating cells, are a small, specialized malignant cell subpopulation with advanced properties of self-renewal, metastasis, and treatment tolerance.^30,42 The CSC theory is supported by increasing evidence.^7,31 CSCs were first described in acute myeloid leukemia and have since been isolated from other hematologic malignancies and many solid cancer types.^31,43,44 However, the origin of CSCs is not clearly understood. Currently, there are two main hypotheses regarding this topic. One hypothesis is that CSCs originate from differentiated non-stem cells that acquire stem cell-like features after malignant transformation.⁴⁵ EMT plays a major role in this process.³¹ Activating an EMT signature could induce numerous stem cell features, including self-renewal capacity and stem cell-related biomarker expression.^45,46 The other hypothesis suggests that non-malignant stem cells become CSCs through transformation induced by oncogenic somatic mutations.^30,31

Our study shows that cervical CSCs exhibit features of both hypotheses mentioned above.³¹ First, we found that cervical CSCs originate from non-malignant cervical stem cells in the normal cervix. We proposed that non-malignant cervical stem cells develop genomic instability and somatic mutations under the long-term influence of HR-HPV genes. Eventually, non-malignant cervical stem cells transform into CSCs (Figs. 2, 3 and 9). Second, during malignant transformation, functional changes of cervical stem cells were characterized by two major features: original non-malignant stem cell properties characterized by high proliferation and low differentiation gradually diminished; however, cancer stem cell properties characterized by EMT, invasion, and metastasis progressively increased (Fig. 3). This progression is supported by multiple results. A hypoxia signature was significantly enriched in CSCs of the ICC group. Hypoxia induces EMT and promotes tumor cell invasion and metastasis (Fig. 3e).^28,29,35 ssGSEA analysis of hallmark pathways also revealed that EMT-promoting IL-6, Wnt, and TGF-β pathways were significantly enriched in CSCs of the ICC group (Fig. 3f).^28,29 In addition, enriched glycolysis and OXPHOS pathways related to glucose metabolism revealed a similar pattern. Glycolytic capacity was increased in CSCs of the ICC group, and OXPHOS profiles were stronger in highly proliferative non-malignant cervical stem cells of the CON group (Fig. 3f). Highly proliferative cells have increased OXPHOS capacity, whereas EMT promotes glycolysis that facilitates tumor cell metastasis.^30,42 The changes in EMT score also illustrated gradually enhanced EMT properties of CSCs during transformation (Fig. S4c). These results suggest that EMT plays an important role in cervical stem cell transformation, which might be an important distinguishing feature of CSC against non-malignant stem cells.

Despite the current understanding of HPV-induced oncogenesis in cervical cancer, our knowledge about which cells promote the development of cervical cancer remains limited. The epithelial organization in the cervix is complex and includes at least three different epithelial types: stratified epithelium of the ectocervix, single columnar epithelium of the endocervix, and transformation zone epithelium.^47,48 Most cervical cancers originate in the transformation zone. For many years, it has been hypothesized that cervical cancer originates from stem cell-like cells known as cervical reserve cells. These cells are variably distributed below the columnar epithelium close to the cervical squamous columnar junction (SCJ) and are involved in cervical metaplasia.⁴⁹ However, some studies have suggested that cervical cancer may originate from a discrete population of cuboidal cells located in the SCJ.⁵⁰ These SCJ cells give rise to a reserve cell-like population that is implicated in cervical metaplasia.⁵⁰

We observed biomarker expression in reserve cells⁴⁹ and SCJ cells⁵⁰ in epithelial and tumor cell clusters (Fig. S3a). CSCs exhibited high reserve cell marker expression (MKI67, KRT17, and TP63); however, the distribution of these markers lacked specificity. In contrast, SCJ cell biomarker expression (CK7, AGR2, MMP7, GDA, and CFTR) in CSCs was low (Fig. S3a), suggesting that the long-presumed reserve cells could be the stem cells observed in our study. However, corresponding marker evaluation via immunohistochemistry and bulk RNA-seq lacks specificity. We proposed new markers for cervical stem cells (Table S9) based on our scRNA-seq data. CSCs contribute to tumor progression, therapeutic resistance, and disease recurrence. Therefore, elucidation of the molecular characteristics of cancer stem cells will improve the effectiveness of targeted therapies for this key group of cells. The specific markers of cervical CSCs identified in this study could be used as potential hotspots for targeted therapies. SCJ cells were not identified in our study, likely due to small SCJ cell numbers.

Additionally, our study revealed a highly heterogeneous cellular composition in the cervical microenvironment. The changes in T cell function are characteristic of cellular transformation processes. During cervical neoplastic lesion development, CD8+ T cells gradually transition to highly activated-exhausted dysfunctional state.^37,38 This seemingly paradoxical activation of CD8+ T cells might be associated with chronic HR-HPV infection, which could cause long-lasting inflammatory responses in the cervical microenvironment.

Global upregulation of IFN responses in the cervical multicellular ecosystem is another prevailing feature during lesion progression. IFN responses play an important role in clearing viral infections.^51,52 However, IFN pathways also function as key drivers of inflammation and immunosuppression in chronic viral infections and cancer.⁵² Therefore, we hypothesized that IFN responses are double-edged swords in cervical cancer. Despite initial functions against HPV infection, sustained IFN responses from chronic viral infection induce a dysfunctional immune state that cannot eliminate the disease, thus promoting HPV persistence and tumor development. This feature of the tumor microenvironment has only been reported in EBV-associated nasopharyngeal carcinoma so far.³⁹ In general, our results suggest that HPV infection has a comprehensive impact on tumor microenvironment formation and evolution during cervical cancer. In particular, viral shaping of the tumor microenvironment may be a common feature of virus-associated tumors.

Our study has several limitations. Firstly, we could not collect low-grade squamous intraepithelial lesion (LSIL) samples, and we thus could not include early precancerous lesions. Secondly, the clonal relationship between T and B cells was not examined. Novel scRNA-seq techniques (including B and T cell receptor profiling and spatial transcriptomics) will help to further explore the intrinsic nature of cervical cancer. Thirdly, this study did not include in vivo and in vitro functional validation experiments at the current stage. Our subsequent research will further investigate the cervical microenvironment with function experiments and large clinical samples. In conclusion, we comprehensively investigated the cervical multicellular ecosystem across progressive cervical cancer stages, with a focus on identifying and characterizing cervical CSCs.

Contributors

Tao zhang participated in the conception and design of the study. Tao Zhang and Liang Zhuang carried out the analysis and interpretation of data, and writing of the manuscript. Munawaer Muaibati, Dan Wang, Abuduyilimu Abasi and Qing Tong have been involved in the examination and critical manuscript revisions. Ding Ma, Lei Jin and Xiaoyuan Huang participated in its design and coordination, and helped to draft the manuscript. All authors have read and approved the final version of the manuscript. Tao zhang and Xiaoyuan Huang have directly accessed and verified the underlying data reported in the manuscript. Lei Jin and Xiaoyuan Huang were responsible for the decision to submit the manuscript.

Data sharing statement

All data are available in the main text, the supplementary materials and GSA-Human database (https://ngdc.cncb.ac.cn/gsa-human/; ID: subHRA005765).

Declaration of interests

The authors declare that they have no competing interests.